Tech • AI • Robotics
Washington Hones in on AI Safety, What Are AI Doomers Proposing | Diet TBPN
7/10TL;DR
Prominent AI slowdown advocates are not proposing a blanket ban on artificial intelligence, but a tightly regulated pause on frontier training paired with state-controlled pathways to reach superintelligence more slowly, around 2040 instead of the late 2020s.
KEY POINTS
A slowdown, not a shutdown
The core proposal centers on delaying the most advanced AI development rather than stopping AI use outright. Existing models would continue to be deployed for inference, and research on current systems could still yield practical products and economic gains. The main objective is to avoid a rapid jump to highly autonomous or superhuman systems before governments and institutions are prepared.
Pause frontier training runs
The most concrete demand is an AI pause on new frontier-scale training runs and related experiments. Under this framework, major compute clusters would be restricted to serving existing models rather than training more powerful ones. The target is the next generation of capability jumps, not routine consumer or enterprise applications built on current models.
Data centers above 10,000 H100 equivalents would face oversight
A proposed enforcement trigger is any facility with more than 10,000 H100-equivalent chips, roughly $100 million in hardware. Such sites would likely need permits, inspections and workload verification. The underlying assumption is that frontier AI remains physically concentrated in large, visible, power-hungry data centers that are easier to monitor than diffuse software projects.
Compute inventories would be declared internationally
Major countries would be expected to disclose national AI compute inventories, including how much advanced hardware they possess and where it is located. The model resembles nuclear non-proliferation: transparency, auditing and mutual monitoring. Supporters see that as the only plausible route to avoid a destabilizing race, while critics note that international agreements are difficult to negotiate and even harder to enforce.
New R&D facilities would resemble secure national labs
Future frontier training, if permitted at all, would move into purpose-built facilities with nation-state-level security. Suggested controls include Faraday cage isolation, air-gapped systems, tightly controlled physical access and extensive verification of personnel and hardware. The intent is to make model theft, covert training and unauthorized deployment far harder.
Bandwidth caps aim to prevent model exfiltration
One unusual safeguard is a hard cap of 1 megabit per second on external communications from top-tier R&D centers. Operators could send instructions into a training site, but extracting model weights would become conspicuously slow, potentially taking years. The idea reflects a broader effort to embed safety into infrastructure, not just policy.
Model transfers would require state supervision
When frontier model weights move from an R&D facility to an inference site, one proposal calls for transfer via physical storage devices with encryption approved independently by both the United States and China. Representatives of both sides would oversee the handoff. That provision illustrates how deeply some advocates envision great-power co-management of the most advanced systems.
Compute caps are viewed as the strongest lever
Among all measures, chip controls and a slowdown in data-center buildout are seen as the most effective ways to restrain capability growth. The reasoning is that hardware concentration is more governable than algorithms, which can spread quickly once discovered. Supporters prefer gradual capability gains tied to monitored compute allocations rather than sudden leaps enabled by breakthroughs that are hard to contain.
The timeline aims for expert-level AI by 2035 and superintelligence by 2040
The slowdown agenda still assumes continued progress toward transformative AI. One schedule would stretch development toward roughly top-human-expert capability around 2035, hold there for several years, and then permit superintelligence around 2040. That contrasts sharply with forecasts that place comparable systems in 2027, 2028 or 2029.
Critics warn of authoritarian control and market distortion
Opponents argue that regulating what can be computed on large clusters is illiberal, vulnerable to regulatory capture, and likely to entrench incumbent firms. Smaller competitors could be frozen out if only a handful of approved players are allowed to operate at scale. Others warn that broad criminal penalties, including proposals resembling a corporate death penalty and prison terms of up to 20 years, could create legal uncertainty around what counts as prohibited AI development.
CONCLUSION
The emerging doomer agenda is less about banning AI than about placing the most powerful systems under heavy state supervision, strict compute controls and international monitoring. The central political fight is whether slowing frontier development would reduce existential risk or simply concentrate power while delaying widely expected economic gains.
Explain this